The New York Times published a piece this morning with the number in the headline: $137 Million in A.I.-Generated Ads Flood the Midterms. Stuart Thompson’s analysis, built on tracking from the Wesleyan Media Project, found more than 220 campaign videos this election cycle that used AI-generated imagery or voices, and more than $137 million spent to put them in front of people. Most of that spending is tied to Republican candidates and groups. About seven in ten of the videos never mentioned that they were machine-made.
The example the Times leads with is New York. Bruce Blakeman, the Republican candidate for governor, released an ad in which Mayor Zohran Mamdani and Gov. Kathy Hochul appear in a sunny backyard, hawking a “political delusion pill” for voters worried about affordability and crime. “Side effects may include: chronic wallet shrinkage, higher tax,” a narrator says over plucky guitar. The faces are real officeholders. The voices are theirs. The video is generated, and the only disclosure in it is the small “paid for by” line. Nothing marks the AI.
A comedy bit, sure. It is also a candidate for governor putting scripted words into the mouths of two sitting officeholders and paying to distribute the result.
Two hundred twenty ads
The scale is what makes this a story about more than one campaign. Two years ago an AI-made ad was a stunt people forwarded around. This year the Wesleyan researchers counted more than 220 of them, spanning subtle touch-ups to full deepfakes, and the spending crossed $137 million. The tech stopped being a novelty in a single cycle, because making a synthetic video of your opponent got cheap enough that a PAC can do it by the dozen. The disclosure question didn’t keep up: no law applied in most of the places these ran, and where laws existed, most of the videos ignored them anyway.
The arrangement, automated
Chapter 2 of the book is titled “News for People Like You,” and it describes an older version of this business. A publication builds an audience by covering what that audience cares about. The audience rewards coverage that confirms what it already believes — with subscriptions, with return visits, with engagement — and over time the outlet learns what its readers want to hear. The pressure on the editor never arrives as an order. It’s less space for the story the readers hate, or a slot behind a story they’ll like better, and each decision can look reasonable on its own. But when preserving that reassurance matters more than reporting an unwelcome finding, the chapter says, the relationship “becomes a form of sycophancy.”
The midterm ads are that arrangement with the newsroom removed. The old loop needed an institution: build an audience over years, watch what they click, let the feedback slowly bend the coverage. The new loop needs a budget. The audience research and the finished ad are now the same machine — generate the version of the candidate that the target voters want to be true, test it, regenerate it, buy the airtime. There is no editor in the loop at all, and nothing upstream for a reader’s trust to attach to.
The anger part
Chapter 2 also makes a point people forget about sycophancy: it isn’t only agreement. It’s also learning what makes someone angry, and serving that. News that provoked a reaction traveled further, so publishers began choosing stories because they would provoke one. I write in that chapter about my own years in news, when a comment thread full of fights could matter more to traffic than the importance of the story. The AI ad runs the same loop without anything left to slow it down. A fabricated video of your opponent saying something outrageous is built to travel, and fabrication makes it cheap to produce and hard to fact-check in advance, because the claim isn’t really in it, only the feeling.
So that’s where the season stands: more than 220 videos, most of them never labeled, and a broadcast environment where the honest default assumption is that what you’re watching was generated for an audience like you. The ad isn’t trying to inform you. Keeping you angry at the right people is the product, and this year $137 million went into building that web — the one that agrees with you.
The book this site accompanies is the long version of the argument, and chapter 2 is where the loop gets built: distribution finds its audiences, audiences reward reassurance, and the people who make the coverage learn to supply it. The midterm ads are the same lesson, minus the humans.